Personalized State-Transition-Aware Memory for Clinical Agents

2026-09-29

AI Summary

This study addresses the challenges of state temporal ambiguity and historical evidence overshadowing in clinical agents caused by memory accumulation. To this end, we propose STAM, a large language model-based State Transition-Aware Memory framework that achieves precise memory maintenance by recording state changes and decoupling active from historical memories. Furthermore, it innovatively integrates semantic retrieval with typed clinical relations to identify affected memories, and employs a query-dependent gating mechanism for the selective invocation of historical information. Extensive evaluations across four longitudinal clinical benchmarks demonstrate that the proposed approach significantly improves downstream question-answering accuracy and enhances diagnostic performance in state maintenance.

Abstract

Large language model (LLM) agents that reason over clinical records must track changes in a patient's state while preserving the history needed to understand them. Simply accumulating memories leaves it unclear which information still applies, whereas overwriting earlier memories can erase evidence needed to reconstruct treatment history and clinical trajectories. We introduce STAM, a state-transition-aware memory framework that records state changes as new clinical entries arrive. STAM combines semantic retrieval with typed clinical relations to identify affected memories, maintaining current information in Active and superseded or resolved information in History. At read time, a query-dependent gate selectively serves historical memory. Across four longitudinal clinical benchmarks, we evaluate STAM with downstream question answering, direct state-maintenance diagnostics, and comparisons at approximately matched context lengths.

Problem

• clinical agents
• state-transition tracking
• memory management
• longitudinal clinical records
• large language models

Innovation

• State-Transition-Aware Memory
• Clinical LLM Agents
• Semantic Retrieval
• Query-Dependent Gate
• Longitudinal Clinical Reasoning